System for Fast and Accurate Visual Domain Adaptation

1. A system for fast and accurate visual domain adaptation, comprising: a computer system having a domain database storing images from different domains, a cloud server with the Generative Adversarial Distribution Matching (GADM) algorithm deployed, said computer system trained by accessing the domain database and applying GADM to perform domain adaptation, wherein the GADM is a Generative Adversarial Network (GAN) based visual domain adaptation incorporating Maximum Mean Discrepancy (MMD) based distance constraint in GAN's objection function, wherein the GAN structure utilizes a single GAN framework comprising a single generator, a single domain discriminator, and a single classifier, wherein source feature is kept unchanged and adaptation is conducted towards target data, and wherein the GADM algorithm is composed of three steps: source domain pre-training, adversarial distribution matching across domains, and target domain classification.

2. The system of claim 1 wherein the source domain pre-training is conducted on sufficient labeled images from source domain and then obtains a source classifier by a standard loss function in supervised learning.

3. The system of claim 1 wherein the adversarial distribution matching across domains is based on GAN's original architecture: one generator network, one discriminator network, one classifier network.

4. The system of claim 3 wherein the generator network is input with target domain images and used as a feature extractor for target data, and is designed with an objective function to minimizing the distribution difference between source feature and target one.

5. The system of claim 3 wherein the generator's objective function contains two parts: the domain discrimination loss term and the proposed MMD loss term.

6. The system of claim 3 wherein the discriminator network is input with both source features and target features and is used as a domain classifier with an objective function to minimizing the domain misclassification.

7. The system of claim 3 wherein the classifier network is input with source domain images and is also used as a feature extractor for source data.

8. The system of claim 1 wherein the target domain classification is based on the pre-trained source classifier.